On the use of the convolutional autoencoder for Arabic writer identification using handwritten text fragments

Abstract

Convolutional autoencoders (CAE) are designed to reconstruct the input image to the output in a near-perfect way via a compact data namely encoded data containing relevant features. The encoded data can be used in various applications as for compressing or classifying the image. The present paper tries to investigate the use of the CAE for writer identification using handwritten text fragments. Hence, the CAE is used for generating features, which is fed to the distance-based classifier. Experimental evaluation is performed on the wellknown IFN/ENIT dataset containing 411 writers. During training, a subset is selected from the 411 writers containing only 11 writers allowing to produce a lite CAE. Experimental results show an identification rate of 92.70% using the whole dataset when the feature vector is appropriately normalized.